We are for the singularity

I. The scarcity is intelligence, not matter

Every unsolved problem is a thinking problem. Cheap intelligence touches all of them at once.

Behind every large unsolved problem there is a humanity that could not think hard enough. Cancer is a search problem: the space of possible molecules is far larger than a human lifetime can scan. Clean energy is an engineering problem. Poverty is a problem of production and allocation.

None of these is solved with more steel. All of them need more attention, more trials, more understanding. Until now the only way to multiply those was to raise more people, which is slow, expensive and limited.

For the first time there is another way. Intelligence can now be built, copied and run at the speed of electricity. Multiplying it is the only move that touches every one of these problems at once, because they share the same bottleneck.

Making intelligence abundant is the most generous thing humanity can do for itself.

II. The curve is real, and it is getting steeper

The work an AI can finish alone doubled every seven months. Now every four.

• 7 mo → 4 moDoubling time of the task horizon: 2019–2025, then 2024–2025.

METR measures the length of task, in human-expert time, that frontier AI agents complete on their own. From 2019 to 2025 that length doubled about every seven months. Since 2024 it has doubled roughly every four; on METR's updated 2026 method, every 89 days.

The trend is robust to our own errors: METR notes that if its absolute measurements are off by a factor of ten, that only moves the arrival of a given capability by about two years. A curve that steep does not care much about where exactly you draw it.

What it looks like

When ChatGPT launched in 2022 it could do coding tasks that take a person about 30 seconds. Today's agents autonomously finish coding tasks that take a human over fourteen hours, with, in AI Digest's words, “no evidence of plateauing.”

Sources: METR, Measuring AI Ability to Complete Long Software Tasks (March 2025) and Time Horizon 1.1 (January 2026) · AI Digest, Time horizons · Ord, arXiv:2608.14426

III. The self-improvement loop has already started

Machines have begun to improve the machines. That is where compounding comes from.

• 20% → 50%An agent that rewrote its own code, on the SWE-bench coding benchmark.

The singularity is not about any one smart machine. It is about the loop: intelligence that makes better intelligence, which makes better intelligence faster. The first turns of that loop are now on record.

The Darwin Gödel Machine is an agent that edits its own code. It improved its own tools and workflow and raised its score on SWE-bench from 20.0% to 50.0%, and on Polyglot from 14.2% to 30.7%, with no human redesigning it. Google DeepMind's AlphaEvolve found algorithmic improvements that cut the training time of large language models by 1%. Agents now carry out machine-learning research tasks on their own, and are ranked on public benchmarks for it.

A 1% gain sounds small. It is the first percent of something that compounds.

The last invention humanity needs to make is the one that makes the rest.

Sources: Zhang et al., Darwin Gödel Machine, arXiv:2505.22954 · Whitfill & Wu, Will Compute Bottlenecks Prevent an Intelligence Explosion?, arXiv:2507.23181

IV. Scaling has room to run

Training compute grows fourfold a year, and can grow ten-thousand-fold more by 2030.

• 2e29 FLOPTraining scale Epoch AI finds physically reachable by 2030, about 10,000× today's frontier.

Frontier training compute has grown about four times a year. That is faster than the adoption of mobile phones, faster than solar installation, faster than genome sequencing. Epoch AI modelled the four things that could stop it (power, chips, data and latency) and found none of them binding before 2030.

That leaves room for training runs of about 2e29 FLOP: a jump as large beyond GPT-4 as GPT-4 was beyond GPT-2. The clusters to do it would cost over $100 billion. The largest training runs of 2030 would take the biggest AI cluster of 2020 more than 3,000 years.

Possible is not the same as inevitable. We argue for it because it is worth building.

Sources: Epoch AI, Can AI scaling continue through 2030? (August 2024) · Epoch AI, What will AI look like in 2030? (2025)

V. Abundance is where this goes

When machines can do most work, growth can run ten times faster. That is the end of scarcity.

• $60TPaid for human labour worldwide every year: the scale of what machines that can do it are worth.

Standard economic models say that if AI can do most or all human work, economic growth could accelerate tenfold or more. Since the industrial revolution the world economy has doubled about every fifteen years; Robin Hanson estimates a comparable AI revolution would see it double at least every quarter.

What that means in ordinary terms is abundance. Altman expects intelligence and energy, “ideas, and the ability to make ideas happen”, to become wildly abundant in the 2030s. Amodei describes “the lifting of billions of people out of poverty to share in the new technologies.”

Poverty has always been a shortage of productive capacity. The singularity is the end of that shortage.

Sources: Epoch AI, Can AI scaling continue through 2030? · Erdil & Besiroglu, Explosive growth from AI automation, arXiv:2309.11690 · Hanson, via Wikipedia, Technological singularity · Altman, The Gentle Singularity · Amodei, Machines of Loving Grace

VI. The first gift is health

A century of biology in a decade. Most diseases defeated.

• 85% vs 20%Microsoft's diagnostic system against human doctors on the NEJM's hardest case challenges.

Amodei's argument is the clearest: AI-enabled biology and medicine could compress the progress human biologists would have made over the next 50–100 years into 5–10. He calls it the compressed 21st century, and its fruits are specific: the defeat of most diseases, and a human lifespan that could plausibly double again.

The first signs are already measurable. On the New England Journal of Medicine's case challenges, Microsoft's MAI-DxO reached 85%; human doctors max out at about 20%.

Sources: Amodei, Machines of Loving Grace (October 2024) · Suleyman, Towards Humanist Superintelligence (November 2025)

VII. Then the robots

Minds that can design will run machines that can build. That is the industrial explosion.

• 327 papersMeta-analysis showing robot models follow scaling laws and improve faster with scale than language models.

Intelligence alone changes what we know. Intelligence in a body changes what we have. Once machine minds can direct machine hands, turning an idea into a medicine, a power plant or a home stops waiting on human labour.

The scaling laws that drove language models hold for robots too: a meta-analysis of 327 papers found robotic task performance improves with scale notably faster than language tasks. Altman writes that 2027 may see robots that can do tasks in the real world, and that robots that can build other robots “aren't that far off.” Forethought estimates that once that happens, industrial output could double every few years or even months.

Sources: Sartor & Thompson, Neural Scaling Laws in Robotics, arXiv:2405.14005 · MacAskill & Moorhouse, Preparing for the Intelligence Explosion, arXiv:2506.14863 · Altman, The Gentle Singularity

VIII. Every year of delay has a price

The risk of going fast is always counted. The cost of going slow never is.

• AsymmetryThe risk of accelerating is always counted; the cost of slowing down never is.

The debate is lopsided. The risks of acceleration are listed in detail. The cost of delay is never written down. Yet every year without the compressed 21st century is a year of diseases without cures, poverty without an exit, and science without the minds to do it.

If powerful AI can compress decades of medical progress into years, then a decade of delay is not caution. It is decades of progress deferred, paid for by people who will not live to see it.

Slowing down is not neutral. It is a decision with victims too.

IX. Safety is not the brake. It is the steering.

A system nobody can steer is not powerful. It is only dangerous. Safety is what lets us go fast.

We want the singularity to arrive, which means we want it to arrive well. The first serious accident would cost everyone a decade. Interpretability, testing and clear thresholds are not the opposite of speed; they are what makes speed survivable.

The frontier labs agree on the structure: define dangerous capability thresholds, test for them before crossing, commit in advance to a response. DeepMind's framework now explicitly covers models that could accelerate AI research itself. We support that work and fund open research that makes powerful systems more reliable.

Sources: Anthropic RSP v3.4 · Google DeepMind Frontier Safety Framework v3.1 · Future of Life Institute, Statement on Superintelligence

X. The human is the end, not the instrument

All of this is for people. If it ever stops being for people, we stop.

We want superintelligence because people suffer, die young, and live far below what they could be. That is the whole reason. We are not looking for another.

A world of machines smarter than us is only the best world if it is a world for us: longer lives, freer minds, work chosen rather than forced, and the wonders of the next century arriving in this one.

If the trajectory of this technology ever stops treating the human as the end, everything this manifesto argues for becomes void. We write that sentence here so we can come back and read it.

Sources: Amodei, Machines of Loving Grace · Suleyman, Towards Humanist Superintelligence